Resolution-adaptive generative ensemble prediction with integrated global-to-regional modelling on unstructured meshes
Integrated global-to-regional ensemble prediction represents an important frontier in Earth system science, yet its realization is hindered by the prohibitive computational cost of running large ensembles on the associated unstructured variable-resolution meshes. Here we introduce HiRe-EFM, a generative deep learning model designed for integrated global-to-regional large-ensemble weather prediction. To address the dual challenges of geometric complexity and multi-scale information flow, HiRe-EFM employs a geometrically consistent mapping onto hierarchical meshes, implements adaptive cross-scale fusion, and incorporates localized enhancement to restore details in refined regions. Benchmarked against a state-of-the-art baseline on a global variable-resolution simulation dataset, HiRe-EFM improves global CRPS by 12.8% while maintaining well-calibrated ensemble dispersion. Crucially, the improvement rises to 16.1% within high-resolution target areas, demonstrating HiRe-EFM exploits the advantages of global variable-resolution simulations. By accelerating global ensemble forecasting by over three orders of magnitude, this study provides a viable pathway for integrated global-to-regional large-ensemble weather prediction.
Authors
- Yihui Zhou (ORCID: https://orcid.org/0009-0002-0368-9547)
- Yi Zhang
- Tao Wu
- Wei Sun
- Meixia Ren
- Jian Li
- Binqian Zhang
Publication Details
- Journal
- npj Climate and Atmospheric Science
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1038/s41612-026-01579-2
- Primary Topic
- Meteorological Phenomena and Simulations
- Type
- article
- Field-Weighted Citation Impact
- 0.00